Audit: broken code

SkillDev tools

Audit dimension - find broken code by running every committed verification instrument and tabling deltas against the recorded baselines. Invoked by dr-audit-orchestrator when broken.md is missing.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Audit: broken code skill

What this skill tells your AI

The instructions your AI receives, as published by ahepi/deepreason in .claude/skills/dr-audit-broken/SKILL.md and read by ahel’s review.

Entry: LEDGER.md exists, broken.md missing. Exit: broken.md written, LEDGER rows added, proofs in proof/.

Broken means: an instrument's output moved from docs/AUDIT_BASELINES.md §Instruments. The instruments define broken; this worker adds no judgment of its own (G3).

Operations

  1. pip install -e . --break-system-packages -q and pip install pytest pytest-xdist jsonschema --break-system-packages -q. Run every command below with python -m where applicable; save each full output to proof/broken-<n>.txt.
  2. python -m pytest tests/ -q -n 4 — compare failures to the baseline failure list. Re-run any non-baseline failure serially (python -m pytest <nodeid> -q) before rowing it: still red → verdict broken; green serially → verdict flaky (row it; flaky is a finding unless baseline-listed).
  3. python tools/docs_verify.py — compare failing checks to the baseline list. Non-baseline failure → verdict broken (the check or the code it pins moved; the fix tranche decides which).
  4. python scripts/wheel_smoke.py then python -u scripts/wheel_operational_smoke.py — any non-zero exit → verdict broken, target = the pin named in the output.
  5. RETIRED (operator ruling 2026-08-22, CLAUDE.md §Build and test): the root sweep is no longer an audit instrument. Do NOT run tools/root_sweep.py. Write proof/broken-sweep.txt containing the single line retired 2026-08-22 — see CLAUDE.md so the proof count stays auditable.
  6. Write broken.md: the table of every non-baseline delta, one row per finding, columns as the LEDGER. Matches baseline exactly → the table plus one line: all instruments at baseline, with the five proof files as the required proof of looking (G2).

GATE

Pass: every instrument ran (four live instruments + the retirement marker file = five proof files) AND every non-baseline delta has a LEDGER row with disposition parked. Verdict labels: broken | flaky | baseline.

Activation plant (first run)

Edit one assert in a copy of a passing test to a false constant, run step 2 scoped to that file, paste the red output, git checkout -- the file, paste the clean status.

Outlets

SituationOutlet
Any fix impulsePARK — prompt in PARKED.md, route deepreason-orchestrator
Instrument itself crashesrow verdict broken, target = instrument, PARK
Baseline looks wrongrow it, PARK a baseline-correction prompt (PRECEDENCE 2)

Signals

GitHub stars
142
Forks
14
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
dr-audit-broken
Source
github.com/ahepi/deepreason